An AI-native animated series for pre-school children, in Gujarati
Prospicience original IP
Khushi Balvatika is Prospicience's own Gujarati animated series for pre-school children. Episode one runs about two and a half minutes across 32 shots, with seven characters in one classroom. A locked style block in every prompt, a signature colour per character and stills approved before motion keep every shot looking like the same film.
SpeedFirst rough cut in under a day
- episode length
- about 2 minutes 30
- shots
- 32
- characters
- 7
- sets
- 1 classroom
A Gujarati language circle time scene for pre-school children. Dialogue is synthesised with a locked voice identity per character, which is what keeps a character recognisable across separately generated shots.
The challenge
Generated animation drifts. Characters change face between shots, the room changes colour, and the result reads as a set of clips rather than an episode.
Why it mattered
Young children follow characters by face and colour, so drift breaks the story for exactly this audience. Traditional animation in a smaller language market rarely clears its budget, so a method that holds consistency at generated video cost changes what a Gujarati series can afford.
What we built
- Wrote a production bible with a locked style block and negative block pasted word for word into every prompt.
- Gave every character a signature colour and set a colour script, so a character stays recognisable even when a shot is regenerated.
- Planned 32 shots with second by second direction across one classroom set, split into 21 talking shots and 11 physics shots.
- Produced Gujarati dialogue with a locked voice per character, plus sound effects and music, assembled into a rough cut.
The result
- A rough cut of episode one: about two and a half minutes, 1920 by 1080 at 30fps, in Gujarati.
- A reusable bible and reference library, so the next episode starts from locked characters, voices and set.
- A tested model choice: one video model now handles both talking and physics shots, so the look stays identical across the cut.
Built for trust
Stills are locked before anything is animated. A still is cheap to judge and regenerate while an animated shot is not, so every mistake is caught at the cheapest possible stage. Model choices are tested before they are trusted: one video model was piloted, could not hold reference images in the mode we needed, and was set aside with the reason recorded.
AI at work
AI generates the character stills, the animated shots and the Gujarati voices. A director locks the style, approves every still before it moves, and checks each take against the production bible.
What it produces
- Character sheets and classroom plates
- Approved keyframe stills
- Talking and physics shots
- Gujarati dialogue with a locked voice per character
- Music and sound effects
How it works, day to day
The director writes the script and shot list against the bible. AI turns each approved still into an animated shot with the character's locked voice, a person checks every take for faces, lips and props, and the approved shots are cut together in the edit.
Where it saves time, and gets better with use
- Mistakes are caught at the still, where a fix is cheap, before any motion is generated.
- Characters, voices and the classroom are locked once, so the next episode starts at the script instead of at character design.
- Each shot gets at most two takes, so time and budget go to the shots that need them.
- The first rough cut came together in under a day.
By the numbers
- about 2 minutes 30
- episode length
- 32
- shots
- 7
- characters
- 1 classroom
- sets
- 1920 by 1080, 30fps
- format
- Gujarati
- language
- 21 and 11
- talking and physics shots
The technology
- An animated series for pre-school children in Gujarati
- The same style rules go into every prompt, word for word
- Every character has a signature colour so they stay recognisable
- Still frames are approved before any motion is generated
- Voices, sound effects and music produced alongside the picture
- Seedance
- ElevenLabs
Technical notes
- Locked style block and negative block reproduced verbatim in every generation prompt
- Per character signature colour as a consistency anchor across independently generated shots
- Reference conditioned generation for dialogue shots, a separate path for physical motion shots
- Documented model evaluation, including one model piloted and rejected with the reason recorded
- Gujarati dialogue synthesis with locked voice identities per character
- Quality gates between stills, animation and assembly rather than one review at the end
Questions we get asked
- How do you stop AI animation drifting between shots?
- By pasting the same style and negative instructions into every prompt without variation, giving each character a signature colour, and approving stills before animating. Drift comes from small prompt differences compounding across shots, so removing the variation removes it at the source.
- Why lock still frames before animating?
- Because a still is cheap to reject and an animated shot is not. Getting the frame right first means every fix happens at the cheapest stage, and the director, not the deadline, decides what ships.
- Can AI animation work in a language like Gujarati?
- Yes, and it is one of the strongest cases for it. Traditional animation in a smaller language market rarely clears the budget, and generated production changes that arithmetic. The dialogue uses a locked voice per character, so every character sounds the same from the first shot to the last.
Related systems
An agentic film studio: brief to storyboard to shots to finished cut
Prospicience internal studio
A brief becomes a finished short film, each shot made by the best-suited model.
4providers routed
First finished cut in under two hours
Media Studio: podcasts, dubbing and video from one brief
Prospicience product for creators and media teams
One brief in, a finished podcast, dub or video out.
10+AI services connected
Which video model should you trust? A vendor-neutral benchmark
Internal research
Six video models, one identical input, scored on what breaks in production.
6provider and model combinations
Run and reported in a single day